I have not yet seen a machine learning algorithm with 20,000 features, but theoretically there is no hard limit. Make sure that you allocate this array on the heap, not on the stack, and that you use an objective and not the trade result for prediction. If this does not help, you can contact Support with that script and they'll look into it.

For a sliding window, I would use either a LSTM neural net, or indeed store the data on the R side. In the prediction part, use an array for storing the features, and shift it by 163 with any new incoming feature set.